The AI Agent is under continuous development. Please see below some current limitations.
Channels
No responses on Booking.com. Booking.com messages still need to be handled by your team. Every other supported channel (Email, SMS, WhatsApp, and Expedia) is covered.
What the agent cannot read in a conversation
These limits apply to what a guest sends the agent during a conversation. Text is the only input the agent processes in a live thread.
No image understanding. A guest sending a photo (of a damaged item, a receipt, a passport) escalates to a human.
No voice notes. Voice messages cannot be transcribed or interpreted by the agent.
No document attachments. A PDF, Word file or scanned paper sent by a guest escalates rather than being read.
This is not the same as importing a document. You can upload a PDF, DOCX, PPTX, HTML or TXT file into the agent's Knowledge tab and have its content extracted into knowledge items — that is how you teach the agent. What the agent cannot do is read a file a guest attaches mid-conversation. See Importing knowledge into your AI Agent.
What the agent cannot do
No PMS actions. It cannot modify a reservation, post a charge, change room assignments, or take any other action in your PMS.
No actions outside the ticket. The agent's reach is the conversation thread. It cannot trigger workflows in other systems unless those workflows are set up to listen to inbox events.
Limits on reservation and guest data
The agent can now answer questions about a guest's own reservation, and it adapts its replies to the guest it is talking to. Both come with boundaries.
Reservation questions
Only on WhatsApp, SMS, and email. These are the channels that can prove who the sender is. Expedia, web messenger, and social channels cannot verify a guest, so the agent redirects reservation questions rather than answering them.
The booking contact must be current. A guest writing from a phone number or email address that is not stored on the reservation cannot verify, even when the booking is genuinely theirs.
Some older email conversations cannot verify. Depending on how the conversation was created, an email thread may not carry the sender's address in a way the system can check. Those conversations hand off to a human. This is a known gap.
Group bookings are not supported. The current design handles one reservation per verification.
For how verification works and how to troubleshoot it, see Answering reservation questions: how guest verification works.
Recognising the guest
Guests must be clustered. Only guests with a master profile get background context. Brand-new and unclustered guests get none, and receive a normal but less tailored reply.
Tone only, never facts. The background summary shapes how the agent writes. It never supplies a verifiable detail, and never contains exact stay counts, dates, or amounts.
Up to 24 hours stale. A CRM note added today may not be reflected until the cache expires.
Not visible to operators on the thread. The summary is not shown alongside the conversation in your inbox. You can read it while testing, though — switch on Guest profile context in the Preview panel and it is printed into the transcript. See Testing Your AI Agent with the Preview Panel.
For what the agent draws on and how guest privacy is protected, see How the AI Agent recognises returning guests.
What the agent cannot learn
No self-training. The agent does not learn from past conversations on its own. It answers from the knowledge you give it, so a new pattern of guest question needs a new knowledge item before the agent can handle it.
Updating knowledge no longer means typing everything by hand — you can import from your website, from a document, or from another agent — but the decision about what the agent should know is still yours. See How the AI Agent learns.
Need Help?
Please contact us through the 'Talk to Us' option on the left menu in the platform, or through the Bookboost Support email at support@bookboost.io if you have questions or need additional support.